Joe: AMD needs, first of all, dual DDR-I for desktop, because as is, it will be completely blown out of the water on some benchmarks. I don't really see any point in launching Athlon64 in single channel version.
This has been my thought as well, but your post and our recent discussions have made me rethink my view on this matter. One of the fundamental problems I "rediscovered" is that many of the applications / benchmarks where the P4 does well in comparison to the Athlon are SSE2-optimized to at least some extent. So the problem becomes, how do we distinguish increases due to SSE2 from increases due to the higher available memory bandwidth?
The easiest thing to do would seemingly be to take a look at the P4 at different FSB / memory settings. So I dug around for some benchmarks…
They're listed below, but because they take up so much space, I've decided to do the conclusion *first*.
CONCLUSIONS:
Let's start with a quick summary, going from 4x100 MHz to 4x133 MHz for the P4 2.4GHz - an increase of 33%:
3D Rendering: Insignificant (<1% increase)
MP3 Encoding: Insignificant (<1% increase)
MPEG2 Video Encoding: Small (~2% increase)
MPEG4 Video Encoding: Reasonable (~8% increase)
Speech Recognition: Reasonable (~7% increase)
Multimedia Benches: Insignificant (<1% increase)
WinRAR: Significant (14.5% increase)
Games are somewhat harder to classify. Basically, you can get a significant increase if you run an "old" game at the lowest resolution without anti-aliasing (12.5% was the highest increase I could find). If you take newer games or run them at higher resolutions the increases quickly falls off, with the video card becoming the bottle neck. With the latest and greatest graphic cards, it is still possible to see reasonable increases (~5%), but only if you stay away from anti-aliasing.
That leaves only the SysMark 2002 Office benchmark which yields a surprisingly large increase of 6.3%.
One very interesting bit in all this is that the benchmarks where the Athlon really gets beat by the P4 (e.g. Lightwave) are completely CPU-bound. I.e. the Athlon gets beaten due to the lack of SSE2.
The Hammer chips will have SSE2. This is very important, IMHO. All indications are that Hammer will actually be even better at executing SSE2 than P4, given its fundamentally better FPU! This should mean that in the benchmarks like Lightwave, Hammer should completely turn the tables on P4 and if not dominate, then at least beat it substantially.
It's also important to remember that the effective bandwidth available to a single channel Hammer will be greater than that of an AthlonXP (using memory of same speed). This is a logical consequence of the lower latency that is attained by having an integrated memory controller. This should translate into substantial gains in most of the benchmarks where memory performance has any effect.
In other words, I don't see a problem for even a single channel Hammer to compete with P4 on 800MHz FSB. Certainly, there are benchmarks where the single channel Hammer will lose, but there's no reason to expect the margins to be anything like what they are with 3D rendering today (due to SSE2 optimizations).
DATA:
Going from 4x100 MHz to 4x133 MHz for the P4 2.4GHz - an increase of 33% (note that I've taken the inverse of the scores given in time to get a consistent sign and comparison ability):Resolution Scores Increase
Quake3 Demo001 (1)
640x480 280.0 fps --> 315.0 fps +12.5% (GF4 Ti4600)
1024x768 275.7 fps --> 307.0 fps +11.4% (GF4 Ti4600)
1280x1024 254.3 fps --> 272.8 fps +7.3% (GF4 Ti4600)
Code Creatures Benchmark Pro (2)
1024x768 29.1 fps --> 29.4 fps +1.0% (GF4 Ti4600)
3DMark 2001 SE (2,1)
1024x768 10783 pts --> 10999 pts +2.0% (GF4 Ti4600)
1024x768 13484 pts --> 14484 pts +7.4% (ATI 9700 Pro)
Serious Sam SE (2)
1024x768 85.4 fps --> 88.8 fps +4.0% (GF4 Ti4600)
Commanche 4 (2,1)
1024x768 46.04 fps --> 46.93 fps +1.9% (GF4 Ti4600)
1024x768 47.13 fps --> 50.03 fps +6.2% (ATI 9700 Pro)
Unreal Tournament 2003 (1)
1024x768 172.7 fps --> 186.3 fps +7.9% (ATI 9700 Pro)
POV-Ray rendering - chess2.pov (2)
n/a 7.818 fph --> 7.852 fph +0.4%
Newtek Lightwave 7.0b (SSE2-optimized) (2)
raytrace scene 28.41 fph --> 28.37 fph -0.1%
reflective_rad 57.1 fph --> 57.5 fph +0.7%
Newtek Lightwave 7.5 (1)
n/a 15.63 fph --> 15.72 +0.6%
Cinema 4D XL 8.001 (1)
Dragon Scene 25.4 fph --> 25.5 fph +0.7%
3D Studio Max 5.1 (1)
Dragon_Charater 30.0 fph --> 30.0 fph 0%
mp3 Maker Platinum (1)
n/a 37.5 1/h --> 37.5 1/h 0%
Lame mp3 Encoding (2)
n/a 35.6 1/h --> 35.6 1/h 0%
Xmpeg DivX Encoding (2)
n/a 13.8 1/h --> 14.9 1/h +8.0%
Pinnacle Studio 8.3.18 MPEG-2 Rendering (1)
n/a 12.00 1/h --> 12.18 1/h +1.5%
Main Concept 1.3 DV to MPEG2 Conversion (1)
n/a 7.993 1/h --> 8.167 1/h +2.2%
PC Mark 2002 MultiMedia Performance (1)
n/a 5891 pts --> 5931 pts +0.7%
Sphinx Speech Recognition
[real / recog] 0.996 --> 1.064 +6.8%
SiSoft Sandra 2003 (1)
Drystones 3147 pts --> 3148 pts 0%
Whetstones 6259 pts --> 6771 pts +8.2%
Multimedia INT 12164 pts --> 12189 pts +0.2%
Multimedia FP 9591 pts --> 9607 pts +0.2%
Winstone 2001 (2)
Content Creation 39.0 pts --> 39.6 pts +1.5%
Business 65.2 pts --> 66.1 pts +1.4%
SysMark 2002 (1)
ContentCreation 330 pts --> 340 pts +3.0%
Office 174 pts --> 185 pts +6.3%
WinRAR
n/a 50.7 --> 58.1 +14.6%
(1) www17.tomshardware.com
These results apply to the NVIDIA Q3 demo as well, but I've been unable to find "crusher" type comparisons. When using Anti-Aliasing (the above scores don't) or higher resolutions like 1600x1200, the increase will quickly become negligible - at least with the current generation of graphics cards (GeForceFX and ATI 9700 Pro).
(2) tech-report.com
-fyo |
